tltorch._tensor_lasso.TuckerL1Regularizer

class tltorch._tensor_lasso.TuckerL1Regularizer(penalty=0.01, clamp_weights=True, threshold=1e-06, normalize_loss=True)[source]

Decomposition Hook for Tensor Lasso on Tucker tensors

Applies a generalized Lasso (l1 regularization) on the tensor layers the regularization it is applied to.

Parameters:	penaltyfloat, default is 0.01 scaling factor for the loss clamp_weightsbool, default is True if True, the lasso weights are clamp between -1 and 1 thresholdfloat, default is 1e-6 if a lasso weight is lower than the set threshold, it is set to 0 normalize_lossbool, default is True If True, the loss will be between 0 and 1. Otherwise, the raw sum of absolute weights will be returned.

Examples

First you need to create an instance of the regularizer:

>>> regularizer = TuckerL1Regularizer(penalty=penalty)

You can apply the regularizer to one or several layers:

>>> trl = TRL((5, 5), (5, 5), rank='same')
>>> trl2 = TRL((5, 5), (2, ), rank='same')
>>> regularizer.apply(trl)
>>> regularizer.apply(trl2)

The lasso is automatically applied:

>>> x = trl(x)
>>> pred = trl2(x)
>>> loss = your_loss_function(pred)

Add the Lasso loss:

>>> loss = loss + regularizer.loss

You can now backpropagate through your loss as usual:

>>> loss.backwards()

After you finish updating the weights, don’t forget to reset the regularizer, otherwise it will keep accumulating values!

>>> loss.reset()

You can also remove the regularizer with regularizer.remove(trl).

Attributes:	loss Returns the current Lasso (l1) loss for the layers that have been called so far.

Methods

__call__(module, tucker_tensor)	Call self as a function.
apply(module)	Apply an instance of the L1Regularizer to a tensor module
apply_lasso(tucker_tensor, lasso_weights)	Applies the lasso to a decomposed tensor
remove(module)	Remove the Regularization from a module.
reset()	Reset the loss, should be called at the end of each iteration.

reset()[source]

Reset the loss, should be called at the end of each iteration.

property loss

Returns the current Lasso (l1) loss for the layers that have been called so far.

Returns:	float l1 regularization on the tensor layers the regularization has been applied to.

apply_lasso(tucker_tensor, lasso_weights)[source]

Applies the lasso to a decomposed tensor

apply(module)[source]

Apply an instance of the L1Regularizer to a tensor module

Parameters:	moduleTensorModule module on which to add the regularization
Returns:	TensorModule (with Regularization hook)

remove(module)[source]

Remove the Regularization from a module.